MosaicML Composer

State of the art algorithms to train your neural networks

Composer is a library for training neural networks better, faster, and cheaper. It contains many state-of-the-art methods for accelerating neural network training and improving generalization, along with an optional Trainer API that makes composing many different enhancements easy.

W&B provides a lightweight wrapper for logging your ML experiments. But you don’t need to combine the two yourself: W&B is incorporated directly into the Composer library via the WandBLogger.

Start logging to W&B

from composer import Trainer
from composer.loggers import WandBLogger

trainer = Trainer(..., logger=WandBLogger())

Use Composer’s WandBLogger

The Composer library uses WandBLogger class in the Trainer to log metrics to W&B. It is as simple as instantiating the logger and passing it to the Trainer.

wandb_logger = WandBLogger(project="gpt-5", log_artifacts=True)
trainer = Trainer(logger=wandb_logger)

Logger arguments

Below the parameters for WandbLogger, see the Composer documentation for a full list and description.

Parameter	Description
project	W&B Project name (str, optional)
group	W&B group name (str, optional)
name	W&B Run name. If not specified, the State.run_name is used (str, optional)
entity	W&B entity name, such as your username or W&B Team name (str, optional)
tags	W&B tags (List[str], optional)
log_artifacts	Whether to log checkpoints to wandb, default: false (bool, optional)
rank_zero_only	Whether to log only on the rank-zero process. When logging artifacts, it is highly recommended to log on all ranks. Artifacts from ranks ≥1 are not stored, which may discard pertinent information. For example, when using Deepspeed ZeRO, it would be impossible to restore from checkpoints without artifacts from all ranks, default: True (bool, optional)
init_kwargs	Params to pass to wandb.init() such as your wandb config etc. See the wandb.init() parameters for parameters that wandb.init() accepts.

A typical usage would be:

init_kwargs = {"notes":"Testing higher learning rate in this experiment", 
               "config":{"arch":"Llama",
                         "use_mixed_precision":True
                         }
               }

wandb_logger = WandBLogger(log_artifacts=True, init_kwargs=init_kwargs)

Log prediction samples

You can use Composer’s Callbacks system to control when you log to W&B via the WandBLogger, in this example a sample of the validation images and predictions is logged:

import wandb
from composer import Callback, State, Logger

class LogPredictions(Callback):
    def __init__(self, num_samples=100, seed=1234):
        super().__init__()
        self.num_samples = num_samples
        self.data = []
        
    def eval_batch_end(self, state: State, logger: Logger):
        """Compute predictions per batch and stores them on self.data"""
        
        if state.timer.epoch == state.max_duration: #on last val epoch
            if len(self.data) < self.num_samples:
                n = self.num_samples
                x, y = state.batch_pair
                outputs = state.outputs.argmax(-1)
                data = [[wandb.Image(x_i), y_i, y_pred] for x_i, y_i, y_pred in list(zip(x[:n], y[:n], outputs[:n]))]
                self.data += data
            
    def eval_end(self, state: State, logger: Logger):
        with wandb.init() as run:
            "Create a wandb.Table and logs it"
            columns = ['image', 'ground truth', 'prediction']
            table = wandb.Table(columns=columns, data=self.data[:self.num_samples])
            run.log({'sample_table':table}, step=int(state.timer.batch))         
...

trainer = Trainer(
    ...
    loggers=[WandBLogger()],
    callbacks=[LogPredictions()]
)